AI4Manufacturing/ITSC-perception

Roles

Roles: perception view of ITSC — annot is the source label (normal / phase_A / phase_B / phase_C), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the repo ships an induction motor's stator-current magnitude in four image encodings as four equal-sized configs — reshaped (consecutive samples arranged as the rows of a grayscale square), scalogram (a continuous-wavelet time-scale view), spectrogram (a short-time Fourier transform) and waveform (the time-domain trace) — 189 records each, and the query states which encoding the record shows. The reasoning column is empty on all 756 records, so this repo is not itself a training view for chain-of-thought and trains annot directly. metadata is provenance (split keys, source identities) and must never be fed to a model.

ITSC — stator fault, perception representations (grounding track)

Part of the AI4Manufacturing FORGE corpus (Category C, task T-C1). Four image renderings of the Park modulus |i_dq| — one channel derived from all three phase currents, so a single-channel view still carries the asymmetry (a balanced set gives a constant modulus; an unbalanced one ripples at twice the line frequency).

Records: 756 across 4 configs (189 windows each); labels {'phase_C': 60, 'phase_B': 57, 'phase_A': 57, 'normal': 15}.

config	image
spectrogram	short-time Fourier transform (time × frequency)
scalogram	Morlet continuous-wavelet transform (time × scale)
waveform	the raw time trace
reshaped	samples arranged into a 2-D grayscale grid

from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/ITSC-perception", "spectrogram")

The reasoning counterpart — the Park-vector locus, which is the representation the physics is read from — lives in AI4Manufacturing/ITSC.

Answer space: use these for is there a fault, not for which phase

annot carries the four-way gold (normal / phase_A / phase_B / phase_C) so the label is not lost, but these four images cannot support the phase, and that is structural rather than a rendering choice. They are drawn from the Park modulus, and the modulus is direction-blind: the shorted phase is encoded in the orientation of the locus, which taking |i_dq| discards. On the shipped features, predicting the phase from the major-axis direction gives 99 %; from the ellipticity — the only phase-relevant quantity the modulus retains — it gives 33 %, i.e. chance.

Measured on the images too. A small ViT (0.55 M params, 40 epochs × 3 seeds) trained from scratch on the pixels alone, balanced accuracy on held-out data:

view	fault vs healthy, old settings	new settings	which phase
scalogram	retracted	retracted	retracted
spectrogram	85 %	90 %	33 %
reshaped	71 %	75 %	24 %
waveform	73 %	73 % (settings unchanged — a control)	47 %
the locus, for contrast	—	—	68 %

The scalogram row is retracted. All three of its numbers were measured on scalogram images whose frequency axis was drawn wrong — the heatmap ran upside-down and the tick labels did not match the rows beside them (present from the tool's first version, fixed 2026-09-17). The ViT was trained on exactly those pixels, so the three figures are void. They have not been re-measured — that needs the scalograms re-rendered first. The other rows come from tools the fix did not touch and stand as they are.

Chance is 50 % / 33 %; the physics reference is 100 % / 99 %. The phase column is measured with severities held out (train 10 %/20 %, test 30 %/40 %) rather than with the shipped repetition split: all twelve conditions appear on both sides of a repetition split, so a model can score by recognising the condition instead of the phase — which inflates waveform from 47 % to 72 % (the scalogram figure that stood beside it is retracted with the row above). Under the honest split spectrogram, reshaped and waveform all fall to chance while the locus does not; whether scalogram does too is now an open question — its figure went with the rest of that row and has not been re-measured.

So: train and report binary on this track, and take the phase from the locus in the reasoning repo. The four-way gold is kept as metadata.

On the render settings. The toolkit defaults were tuned for broadband bearing vibration; on a narrowband 1 kHz Park modulus they gave a spectrogram with only 4 time columns, a scalogram axis pinned to fs/128…fs/2 rather than to any chosen band, and a reshaped image that was 88 % zero padding (500 samples padded to 64²). All are now set explicitly — see _provenance.json → generation.perception_render. waveform needed no change (500 points on an ~880 px canvas), which makes it a useful control: it returns the same 73 % under both, so the differences elsewhere are real rather than run-to-run noise.

One weakness to keep in mind: the dataset holds only five healthy records, two of which land in test — so the healthy side of every number above rests on 12 windows and moves in steps of 8 points.

Rig

Baldor CM3542 three-phase squirrel-cage induction motor, 0.75 hp, 208-230/460 VAC, 1725 rpm at 60 Hz, 59 turns per pole, double-star. Inter-turn short circuits are seeded in one phase at a time at four severities; every measurement is a steady state without load, fed directly from the mains. That last detail is what makes the fault observable in the current at all — on an inverter-fed machine the drive's current controller regulates the asymmetry away.

Schema (7-field unified record)

field	meaning
query	the classification instruction (one of 30 deterministic paraphrases per representation)
image	the rendered signal image (bytes embedded)
annot	gold class: normal / phase_A / phase_B / phase_C
reasoning	chain-of-thought (empty here; filled in the ITSC-annotated sibling)
cate / task	C / T-C1 (signal fault classification)
metadata	JSON string: representation, ellipticity/major-axis/negative-sequence features, Park-modulus time stats, per-phase RMS, the calibrated thresholds, computed_verdict, evidence_tier, fault_phase, severity_pct, line_hz, fs, file, window_idx, image_sha256, split

Splits

train / test = repetitions 1-3 / 4-5 (leakage-safe).

Provenance & the evidence gate

Generated deterministically by forge_agent/examples/itsc/convert.py (4c4873f9c7) → forge_model/ITSC/convert_itsc.py (41f90a674e); see provenance.json for the full record.

Labels come from the dataset's own filenames. An evidence gate — a label-independent computation on the raw current — decides what ships: park_vector_negseq measures the ellipticity of the three-phase current locus and the direction of its major axis, and a record is confirmed only when that blind computation lands on the gold class.

Both thresholds are calibrated on the train split's healthy records only and then applied blind: the ellipticity cut is mean + 3·std = 8.48% (healthy 4.89 ± 1.19%, n=9 windows — thin, and stated as such below), and the phase-axis centres are the circular means of the major-axis angle per faulted phase (A=140.0°, B=89.5°, C=27.4°). No threshold was adjusted against the faulted classes.

A third check verifies Kirchhoff's law on each window (ia+ib+ic ≈ 0) against a fixed, physics-derived cut: inverting one current clamp doubles that residual, and every healthy rig measured sits far below the cut. A reversed clamp would otherwise read as a large false asymmetry — which is exactly what happens on the KAIST PMSM set. Here 0 of 195 windows are flagged.

This track keeps confirmed + weak-nonconflict — a window is dropped only when the blind computation points at a different fault than the gold, never merely for being quiet.

Caveats

• The evidence gate is measured on the SIGNAL, not on the image. park_vector_negseq reads the raw three-phase current; it never opens the PNG. So evidence_tier says this signal supports the label — not this image shows it. Every record carries evidence_measured_on: "signal" so the claim travels with the data. Gating this way is deliberate: the computation is deterministic and will give the same answer in a year, whereas filtering by what a model can already read would select for what it already knows. It is also sound but incomplete — the image is a lossy function of the signal, so no-evidence-in-signal does imply no-evidence-in-image, but a record can pass and still lose its evidence in rendering. Image-side readability has never been measured, here or anywhere in this corpus. When it is, the rule is: evidence missing from the signal → drop the record; evidence present but unreadable in the image → change the rendering, not the record set.
• **Severity is metadata, not the answer.**The source seeds four levels (10/20/30/40%) and they are carried per record, but adjacent levels overlap badly in every measured indicator, so a record cannot be graded to its exact percentage from this evidence. The answer space is therefore presence + location (normal / phase_A / phase_B / phase_C).
• Source variant matters. This build uses the authors' own per-record crop (Cropped_Signals_SF/), not the raw 5 s recordings. In the raw files the short circuit is switched in and out mid-record — the first and last second carry no fault — so windowing them would inherit a record label onto fault-free windows. Measured: 41% of faulted raw windows fall inside the healthy range, against 3% on the authors' crop.
• One crop residue. In SC_A4_B0_C0_004 window 2 the short is switched out 0.22 s into the window: the Park modulus' oscillation collapses from 52% of its mean to 8.7% (ellipse -> near-circle) and the last ~36% of the window is fault-free. It still clears the evidence gate (ellipticity 31.3 vs threshold 8.5) and its label is unaffected, but its pale locus is visibly fanned, and it is the only faulted record whose ellipticity varies more than 20% across its own windows (24.9%; next worst 9.6%, median 1.8%). Left in deliberately — the label is the authors' own and it is correct.
• Small, single-rig source. 65 recordings from one 0.75 hp motor at one operating point (steady state, no load). The value is a groundable stator-fault benchmark, not record volume.
• Ellipticity vs negative sequence — they are the SAME measurement here, and the choice is purely about the image. On the 195 published windows the two are almost perfectly collinear (Pearson r = 0.997; neg-seq ≈ 0.60 × ellipticity, ratio spanning only 0.47–0.68) and they separate identically (each: 6 of 180 faulted inside the healthy range, 98.3 % best balanced accuracy). Do not claim one is more robust than the other — an earlier version of this card said ellipticity is less affected by supply-voltage unbalance; that is not supportable at r = 0.997. Ellipticity is used because it is the quantity the published image actually draws, so a chain-of-thought citing it cites something the reader can see; the negative-sequence ratio has no counterpart in the image.
• No voltage channel. Supply-voltage unbalance also produces an unbalanced current set, and with current-only data it cannot be separated from a winding fault. This limit applies to both indicators equally — it is a property of the dataset, not a reason to prefer one over the other.
• The major-axis angle alone identifies the faulted phase for 95.0 % of faulted windows (cluster means A 139.3° / B 87.0° / C 27.1°). It is only consulted after ellipticity has already said 'faulted' — healthy windows here cluster at 87-117°, overlapping phase B, so the angle is not evidence of a fault on its own.
• Split — train/test follow repetitions 1-3 / 4-5. Each repetition is a separate acquisition and windows never cross one, so it is leakage-safe. It is a repetition split on a single motor at a single operating point, not an unseen-machine split — it says nothing about transfer to another motor.
• The healthy class is thin. Five healthy recordings exist in total, three of them in train, so the ellipticity threshold stands on 9 windows. That is the weakest link in this release.
• Only the healthy-vs-faulted step is label-free. Naming which phase uses the three phase-axis centres, which are the mean major-axis direction per faulted phase over the train split's labels — calibrated there, applied blind to test, but not label-independent. Do not read the whole verdict as label-free.

Source & license

Source: ITSC dataset, Laboratory of Electrical Engineering, Universidad de Guanajuato, Mexico (github.com/ibarram/ITSC). License: MIT. Please cite: Cardenas-Cornejo, Ibarra-Manzano, González-Parada, Castro-Sanchez & Almanza-Ojeda, Classification of inter-turn short-circuit faults in induction motors based on quaternion analysis, Measurement 222 (2023) 113680, doi:10.1016/j.measurement.2023.113680.

Overlap / de-duplication (§8)

Cross-family evaluation lock — metadata.eval_lock (stamped 2026-09-20; manifest revision fe6e286912b0, generated 2026-09-08). Every record of this repository, locked or not, carries metadata.eval_lock, computed by forge_model/common/overlap.py::Overlap.stamp_for against common/overlap_manifest.json at that revision — so within this repository the absence of the key cannot occur. Shape: {"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}. locked is true when the image is evaluation material anywhere in the corpus; against names every repository and split in which it is (sorted; [] when not locked; it includes the record's own family where that is so); own_split marks a record locked by its own split. The per-record field is the authority — the count here is quoted once, at this revision, and a later manifest may change it: 288 of 756 records (288 distinct images) are locked — by column: 0 by the cross-family manifest, 288 by their own split, 0 both ways and counted once; counterparts (records per counterpart; a record can appear under several): none — every lock here is by the record's own split; 288 locked by their own split: test. In words: 288 of the 756 records in this repository are evaluation material by their own metadata.split (test: 288) and sit inside the HF split named test / train — under the uniform-split convention the HF split name is a container name, and metadata.split together with metadata.eval_lock carries the truth; a train pool must exclude them. A stamp whose manifest_revision differs from the current manifest is stale, not wrong — recompute it (Overlap.stamp_is_current); a record with no stamp has not been checked against the corpus as it now is. Overlap.partition / assert_train_pool_clean read the field: a train pool built from this repository must exclude every locked record.